I’m PHAM Thanh Dat

Solution Architecture Software Engineer Computer Vision Engineer AI Researcher

HIRE ME

PHAM

Thanh Dat

ABOUT ME

My name is Dat. A Solution Architecture and Software Engineer based in Hanoi. I have worked on several projects which relate to Computer Vision field based on combining between 2D and 3D images

Also, I have some experiment about Optimise AI Model to speed up the inference time by using TensorRT and CUDA on GPU.

Moreover, as a solution architecture, I have serval experiment of designing the Database and how to optimize the Back-end systems with high frequency request of user.

DOWNLOAD MY CV

EXPERIENCE

Solutions Architect
Asilla - VietNam
01/2024 - Now
– Bring serval features of Asilla product from on-premise to cloud service by using AWS.
– Design serval Cloud Service in order to sync data between cloud and local clients.
– Refactor the code of DeepStream pipeline by using event-drive.
– Optimize the resource and performance of inference AI model by using TritonServer.
– Create the a Infra as Code for easy deployment and maintenance by using Terraform.
AI/Software Engineer
Asilla - VietNam
04/2023 - 12/2023
– Develop some algorithms to detect the abnormal on escalator by using the Pose Estimation as input.
– Improve the performance of system that are implemented based on DeepStream framework using C++.
– Create and maintain serval tools for easy deployment and debug the system in both production and dev environment.
Computer Vision Engineer
Boxy (Storelift) - France
08/2019 - 03/2023
– Create and deploy the tracking system in the autonomous, cashier-less stores.
– Using Object Detection, Background Subtraction, Optical Flow to detect people in the store.
– Using 3D information as Depth, Point Cloud, Calibration to improve the accuracy.
– Optimise and reduce the memory on GPU to adapt the real-time inference with TensorRT, OpenCV, Batch-processing …
– Create and optimize the as small as possible docker size to decrease the time of deployment.
Master 2 Internship
Grenoble Informatics Laboratory - France
01/2019 - 07/2019
– Evaluate the quality of generated data from Generative Adversarial Network(GAN) as the diversity and the similar compared to original data.
– Train a classifier to see the performance of generated images from GAN.
Research Internship
Grenoble Informatics Laboratory - France
01/2018 - 07/2018
– Improving the performance when training deep networks base on multi- GPUs for Coupled Ensembles model

SKILL

Python 90%

C++ 90%

CUDA 70%

Pytorch, OpenCV 85%

ROS2, ZERO-MQ 85%

Tensorflow 50%

Research 80%

DevOps - AWS 80%

Database 70%

EDUCATION

National School of Computer Science and Applied Mathematics of Grenoble
Master 2 of Data Science
2018 - 2019
Hanoi University of Science and Technology
System of Information and Communication
2013 - 2018

FUNFACTS

HANDS TO WORK

CUPS OF COFFEE

Project

Asilla Cloud platform

On the past, Asilla AI Security is on-premise system. By the popular of cloud, Asilla also deploy a cloud service which show serval import information from local to cloud such as alert detection, some analytics … Asilla Cloud platform is additionally a place to setting some feature to on-premise system and sync the results back to cloud. Because the data is updated frequency every 1 minutes, we use some Lambda function to handle the scale and workload. Serval technique about database as partition, caching, optimization query is used to speed-up the performance.
Deployment processing is a main part for cloud platform. In order to reduce the time and mistake by human, all the infrastructure is deployed as Infra As Code by Terragrunt which is a wrapper of Terraform. Some tech stacks are used as auto-scale, web application firewall .. to keep and protect the system alive 24/24h.

Action Recognition on Escalator

Asilla is building an AI security SDK based on AI technologies as Computer Vision, Action Recognition, Pose Estimation … As an AI Engineer, I already developed the algorithm to detect the abnormal action on escalator by using the pose estimation and speed estimation. I also integrate this algo to current Deep Stream Pipeline of SDK by using C++ and Python. The speed estimation also is shared among all downstream plugins in order to reuse at another features of systems. [Source]

Multi-Object Tracking

As Computer Vision Engineer, I create and deploy the tracking and object detection with high accuracy in the real time mode. The algorithms are developed by using the multi-camera. Besides the normal techniques of 2D image processing as Background Subtraction, Optical Flow, Object Detection …, I also use serval algorithms of 3D image as Calibration, Depth and Point Cloud processing to improve the accuracy. Furthermore, I optimise these models for speed and reduce the memory of GPU to adapt the real time inference by using Tensorrt, Opencv with Cuda, Batch processing, Gstreamer, multi-stream in GPU and multi-processing in CPU. To deploy the tracking systems to production, I also create and optimise the Docker file with as small as much the size.
For result, the tracking system is deployed in around 25 stores with around 1000 clients per day. The error always be less than 3%

Monocular Depth Estimation

Nowadays, with the rapid development of technology of cameras, beside of taking the images as 100 years ago, the new cameras can also estimate the depth value (distance relative to the camera) of each pixel in an image like the eyes of human. Combining both of them (Depth and RGB image), it improves and reaches the state-of-the-art of some tasks as Object Detection, Object Tracking in the domain Self-driving car, AR. However, the prices is much higher than the normal cameras. It also require some special setups as a USB camera, hardware. Then, the objective of project is to try estimate the depth value given a single (monocular) RGB image. It means that, we can apply the new algorithms in level of 3D scene with the normal camera to improve the performance of system without adding any hardware and the cost.

Classifier Training from a Generative Model

In this period, we used the generative adversarial networks (GAN) as a method to create data. To evaluate the quality of generated data from GAN, we trained two classifiers. The first one is with the real data, the another one is with the generated data. From the error of validation test, we can see that how the diversity of GAN are
For more details, see at link

CONTACT

I am always looking for good opportunities and new challenges! Feel free to contact me!

Contact me!

I’m PHAM Thanh Dat

ABOUT ME

EXPERIENCE

Solutions Architect

AI/Software Engineer

Computer Vision Engineer

Master 2 Internship

Research Internship